Optimizing ISAAC ROS ESS Depth Estimation to run efficiently with lower GPU memory utilization

My current setup:
Model: Light ESS
Stereo Camera: Intel Realsense 435f

Concerns & Support:
The model itself once loaded utilizes ~2GB of GPU memory without any preprocessing steps. This is turning out to be a major hindrance for us, as we want to run 2 cameras at least at any instance of time for navigating safely on our robot.
With preprocessing steps added for mono8->RGB and resizing we are ending up with ~4.5GB of GPU memory used by a single program. Our current robot uses a RTX 2060 with 6GB of memory which is a major roadblock for us to run this package with two cameras.

I would really appreciate any suggestions or feedback on how we can optimize this model to run with our current compute for 2 cameras as efficiently as possible.

Hi @vishrut1

We’re working to further optimize ESS light for memory footprint without compromising performance

We’re also looking at tuning memory pools across packages to reduce the high watermark. The 2.5GB of additional processing is an issue.

Raffaello

2 Likes